An Introduction to

Interaction Analysis
Tyler J. VanderWeele
Departments of Epidemiology and Biostatistics
Harvard School of Public Health

Overview
Introductory Examples
Additive and Multiplicative Interaction
Statistical Interaction
Mechanistic Interaction
Epistasis in Genetics
Concluding Remarks

Introductory Examples
In many settings, the effect of one exposure may depend in some way
on the presence or absence of another exposure
We then say that there is interaction between the two exposures
Recent years have seen increasing interest in interaction between
genetic and environmental exposures
Interaction can also occur between two (or more) environmental
exposures, or two genetic exposures, or with various behavioral
exposures, etc.
The processes giving rise to illness and health is often inherently
complex
Interaction between exposures is one manifestation of this complexity

Introductory Examples
Figueiredo et al. (2004) studied the effects of XRCC3-T241M polymorphisms and
various environmental factors on breast cancer risk
For XRCC3-T241M using a case-control study they found the OR for breast
cancer for the M/M genotype was 1.47 (CI: 1.00, 2.15) times that of the reference
T/T or T/M genotype
However the effect varied by strata of alcohol consumption:
OR for Breast Cancer (by strata of alcohol consumption and XRCC3-T241M)
T/T or T/M
M/M

No Alcohol
1.00
1.21 (0.70-2.09)

Alcohol
1.12 (0.81-1.54)
2.09 (1.16-3.78)

It seems as though XRCC3-T241M polymorphisms do not have much of an effect

unless accompanied by alcohol consumption
This is an example of what we might call a gene-environment interaction

Introductory Examples
In a review article, Hunter (2005) lists numerous examples of other
gene-environment interactions including
Polymorphisms of:

Environmental Exposure

Outcome

MTHFR
NAT2
APOE
ADH1C
PPARG2

Folic acid

Colorectal cancer
Colorectal cancer
Serum cholesterol
MI
Obesity

Heterocyclic amines in cooked meat

Dietary cholesterol
Alcohol intake
Dietary fat

Likewise we might also be interested in the interactions of two genetic

factors on various outcomes (though there are fewer clearer examples of
this)
We will consider (i) statistical ideas to conceptualize interactions and (ii)
how these relate to more mechanistic or causal notions of interaction

Notation
We will let G denote our genetic factor of interest
We will let E denote our environmental factor of interest
We will let D denote our outcome of interest
For simplicity we will assume that G and E are binary
i.e. E = 0 for the environmental exposure absent; E = 1 for present
i.e. G = 0 for low genetic risk; G = 1 for high genetic risk
e.g. for a genetic factor with dominant mode of inheritance we let:
G = 0 for a/a genotype
G = 1 for a/A or A/A genotype
(or for recessive inheritance G = 0 for a/a or a/A; G = 1 for A/A)
The ideas presented here apply more generally to exposures that
are not binary

Additive Interactions
How do we measure interaction?
Suppose we had the following data on risks from a cohort study:

G=0
G=1

E=0
0.02
0.04

E=1
0.05
0.15

Let pij = P(D=1 | G=i, E=j ).

A natural way to assess interactions is to measure the extent to
which the effect of the two factors together exceeds the effect of
each considered individually:
(p11 - p00) - [(p10 - p00) + (p01 - p00)] = p11 - p10 - p01 + p00
This is sometimes referred to as a measure of interaction on the
additive scale

Additive Interactions
Data:
G=0
G=1

E=0
0.02
0.04

E=1
0.05
0.15

Additive measure of interaction: p11 - p10 - p01 + p00

If p11 - p10 - p01 + p00 > 0 the interaction is said to be positive or
super-additive
If p11 - p10 - p01 + p00 < 0 the interaction is said to be negative or
sub-additive
Here we have:
p11 - p10 - p01 + p00 = 0.15 - 0.04 - 0.05 + 0.02 = 0.08 > 0
i.e. a positive interaction

Multiplicative Interactions
Data:
G=0
G=1

E=0
0.02
0.04

E=1
0.05
0.15

As an alternative to assessing interactions on the additive scale

using risks we might consider a multiplicative scale using relative
risks:
Let RR11 = p11/p00 = 0.15/0.02 = 7.5
Let RR10 = p10/p00 = 0.04/0.02 = 2
Let RR01 = p01/p00 = 0.05/0.02 = 2.5
A measure of multiplicative interaction for risk ratios is:
RR11 / (RR10 x RR01) = 7.5 / (2 x 2.5 ) = 7.5 / 5 = 1.5
If the multiplicative interaction is > 1 it is positive, < 1 it is negative

Multiplicative Interactions
Data:
G=0
G=1

E=0
0.02
0.04

E=1
0.05
0.15

In case control studies it is not in general possible to estimate risks

or even risk ratios but one can still estimate odds ratios:
Let OR11 = {p11/(1-p11)} / {p00/(1-p00)}
Let OR10 = {p10/(1-p10)} / {p00/(1-p00)}
Let OR01 = {p01/(1-p01)} / {p00/(1-p00)}
Interaction on the odds ratio scale is then measured by:
OR11 / (OR10 x OR01)
Again if this is > 1 the interaction is positive, if < 1 then negative
If the outcome is rare then the OR interaction measure
approximates the RR interaction measure

Additive vs. Multiplicative Interactions

Conceived of in this way, interaction depends on the scale
(multiplicative or additive)
We may in fact have additive interaction w/o multiplicative
interaction
E=0
E=1
G=0
0.02
0.05
G=1
0.04
0.10
Additive:
p11 - p10 - p01 + p00 = 0.10 - 0.04 - 0.05 + 0.02 = 0.03 > 0
Multiplicative:
RR11 / (RR10 x RR01) = 5 / (2 x 2.5) = 1

Additive vs. Multiplicative Interactions

In other settings we may have multiplicative interaction but no
additive interaction:

G=0
G=1

E=0
0.02
0.07

E=1
0.05
0.10

Additive:
p11 - p10 - p01 + p00 = 0.10 - 0.07 - 0.05 + 0.02 = 0
Multiplicative:
RR11 / (RR10 x RR01) = 5 / (3.5 x 2.5) = 0.57 < 1

Additive vs. Multiplicative Interactions

For some time in the epidemiologic literature, there was debate as to
which scale one should assess interactions on (Blot and Day,
1979; Saracci, 1980; Rothman et al., 1980)
The general historical consensus was:
(1)Often the additive scale is of greatest public health importance
It allows one to discern whether the effect would be different
in different subgroups (Rothman et al., 1980)
(2)The additive scale also seemed to correspond to the more
biological notion of synergism as conceived of by
Rothman (1976); (later in lecture)
(3)However, sometimes the multiplicative scale (or neither scale)
may be the one that more naturally corresponds to the
biological mechanisms (Siemiatycki and Thomas, 1981);
though in these cases the additive scale is still important for
assessing public health impact

Additive vs. Multiplicative Interactions

Suppose that E denotes some drug and the outcome is cured:
And there are 100 with G=0 and 100 with G=1 and we have 100 doses
E=0
E=1
G=0
0.01
0.05
G=1
0.04
0.10
The risk difference for E on those with G=0 is:
0.05 - 0.01 = 0.04
The risk difference for E on those with G=1 is:
0.10 - 0.04 = 0.06
p11 - p10 - p01 + p00
= 0.10 - 0.04 - 0.05 + 0.01 = 0.02 > 0
The risk ratio for E on those with G=0 is:
0.05 / 0.01 = 5
The risk ratio for E on those with G=1 is:
0.10 / 0.04 = 2.5
RR11 / (RR10 x RR01) = 10 / (5 x 4) = 0.5 < 1
If we give the drug to G=0 group we cure: 100*(0.05) + 100*(0.04) = 9
If we give the drug to G=1 group we cure: 100*(0.01) + 100*(0.10) = 11
We should treat the G=1 group; we cure an additional 2 persons
Additive interaction (not multiplicative interaction) identifies this
The risk ratio suggests treating the G=0 group; for public health purposes we should
rely on additive interaction measure to decide which subgroups to target

Additive vs. Multiplicative Interactions

In case-control studies we cannot in general estimate the risks themselves
and generally odds ratios (which approximate risk ratios when the outcome
is rare) are used.
Assume a rare outcome so ORs and RRs are approximately equal:
We cannot get the additive interaction directly but suppose we divide
Additive interaction p11 - p10 - p01 + p00
We get:
RR11 - RR10 - RR01 + 1

by p00

This gives us something like the additive interaction but using RRs (or
ORs); it is sometimes called the Relative Excess Risk due to Interaction or
RERI (Rothman, 1986)
If RERI > 0 we have a positive additive interaction; if < 0 a negative additive
interaction
We can thus assess additive interaction using risk ratios (or odds ratio if the
outcome is rare)

Additive vs. Multiplicative Interactions

Suppose we had a case-control study
Assume the outcome is rare and we have the following ORs or RRs

G=0
G=1

E=0
1
2

E=1
2.5
5

Although we cannot get at the additive interaction directly we could

estimate:
RR11 - RR10 - RR10 + 1

= 5 - 2 - 2.5 + 1 = 1.5 > 0

Although we have case-control data we know we still have positive

interaction on the additive scale i.e. targeting one group will have
greater public health benefit than targeting another group

Additive vs. Multiplicative Interactions

In most published epidemiologic studies, interactions are evaluated
and reported on the multiplicative scale
Interaction on the additive scale are often not e.g. perhaps only
about 1 in 50 in epidemiology reported (Knol et al., 2009), even
though there has been consensus that it should be reported for
assessing public health relevance
The focus on multiplicative interaction is likely due to the statistical
models which are used in such analyses (e.g. logistic regression)
and the fact that the models employed immediately give
interactions (and confidence intervals) on a multiplicative scale
In general, if interaction are interest it is probably good to report
estimates and confidence intervals on both scales

Statistical Interactions
A statistical model on the linear scale accommodating interaction
takes the form:
P(D=1|G=g,E=e) = 0 + 1g + 2e + 3eg
In the regression setting:
3 = p11 - p10 - p01 + p00
i.e. the interaction contrast on the additive scale
In fact:
0 = p00
1 = p10 - p00
2 = p01 - p00

Statistical Interactions
In the context of the data before:
E=0
E=1
G=0
0.02
0.05
G=1
0.04
0.15
P(D=1|G=g,E=e) = 0 + 1g + 2e + 3eg
We would have:
0 = p00
1 = p10 - p00
2 = p01 - p00
3 = p11 - p10 - p01 + p00

= 0.02
= 0.02
= 0.03
= 0.08

3 is referred to as a statistical interaction on the additive scale

because it arises from a statistical model

Statistical Interactions
Similarly one might have a log-linear model for risk ratios:
log {P(D=1|G=g,E=e)} = 0 + 1g + 2e + 3eg
exp(0) = p00
exp(1)= RR10
exp(3) = RR11 / (RR10 x RR01)

exp(2) = RR01

Or one might alternatively use a logistic model for odds ratios:

logit {P(D=1|G=g,E=e)} = 0 + 1g + 2e + 3eg
exp(0)=p00/(1-p00) (cohort study only) exp(1)=OR10 exp(2) = OR01
exp(3) = OR11 / (OR10 x OR01)
In these cases 3 is again referred to as a statistical interaction but
now on the risk ratio or odds ratio scale

Logistic Regression and the

Delta Method
For RERI, if the outcome is rare (so that odds ratios approximate
risk ratios; or if incidence density sampling is used) we can
calculate RERI using the coefficients of a logistic regression model:
logit {P(D=1|G=g,E=e)} = 0 + 1g + 2e + 3eg
We can then calculate:
RERI OR11 - OR10 - OR01 + 1
= exp(1 + 2 + 3) - exp(1) - exp(2) + 1

Statistical Interactions
Consider again the data from Figueiredo et al. (2004) with odds
ratios compared to the reference category G=0 (i.e. T/T or T/M) and
E=0 (i.e. no alcohol):
T/T or T/M
M/M

No Alcohol
1.00
1.21 (0.70-2.09)

Alcohol
1.12 (0.81-1.54)
2.09 (1.16-3.78)

logit {P(D=1|G=g,E=e)} = 0 + 1g + 2e + 3eg

exp(1) = 1.21
exp(2) = 1.12
exp(3) = OR11 / (OR10 x OR01) = 2.09 / (1.21 x 1.12) = 1.54
There is a positive interaction on the multiplicative scale
Also RERI OR11 - OR10 - OR01 + 1 = 2.09 - 1.21 - 1.12 + 1.00 = 0.76 > 0
There is a positive interaction on the additive scale as well

Statistical Interactions
The software used to fit models such as:
P(D=1|G=g,E=e) = 0 + 1g + 2e + 3eg
logit {P(D=1|G=g,E=e)} = 0 + 1g + 2e + 3eg
will give confidence intervals and p-values for the interaction coefficients,
i.e. 3 and 3 respectively.
These statistical models can also easily accommodate additional
confounding variables or covariates in the model
For RERI, Hosmer and Lemeshow (1992) give standard errors using the
delta method
Lundberg et al. (1996) provides some SAS code to do these computations
automatically
Knol and VanderWeele (2012) provide an easy-to-use Excel spreadsheet
to do this

Case-Only Estimators
One other more recent approach deserves attention
Consider the interaction term 3 in a logistic regression
logit {P(D=1|G=g,E=e)} = 0 + 1g + 2e + 3eg
Suppose also that G is independent of E in the population (plausible in
many gene-environment interaction studies) and that the outcome is rare
Suppose that data are only collected on the cases (D=1)
It turns out that the odds ratio relating G and E among the cases is equal
to the interaction measure on the multiplicative scale exp(3) (Yang et al.
1999; cf. Piergorsch et al., 1994)
P(G=1|E=1,D=1)/(P(G=0|E=1,D=1) = exp(3) = RR11/(RR10 x RR01)
P(G=1|E=0,D=1)/(P(G=0|E=0,D=1)
Essentially to get measures of multiplicative interaction, all that is needed
is data on G and E among the cases

Case-Only Estimators
This is referred to as the case-only estimator of interaction
The case-only estimator depends critically on the assumption of GxE
independence and can be quite biased if this is violated (Albert et al., 2001)
Under this assumption of GxE independence the case-only estimator is in
fact more efficient than using the standard estimate from a logistic regression
We can condition on covariates C to make independence more plausible
With the case-only estimator we can estimate the interaction parameter 3
logit {P(D=1|G=g,E=e,C=c)} = 0 + 1g + 2e + 3eg + 4c
but we cannot estimate the main effect of the logistic regression
Estimates and confidence intervals for the case-only estimator can be
obtained by running a logistic regression of G on E and C among the cases:
logit {P(G=1|E=e,C=c,D=1)} = 0 + 3e + 1c

Mechanistic Interactions?
Do statistical interactions tell us anything about biological or
mechanistic interactions?
Several authors have pointed out the potential danger of using
statistical interaction to draw conclusions about biological
interaction (Siemiatycki and Thomas, 1981; Thomas, 1991;
Rothman and Greenland, 1998; Cordell, 2002)
How might we conceive of mechanistic interaction?
Can we conclude anything about mechanistic interaction from
statistical interaction?

Mechanistic Interaction
Sufficient Causation in Statistics and Epidemiology
Rothman (1976) defined a sufficient cause as minimal set of events,
conditions or characteristics that inevitably produced the disease; a
component cause (or cause) was an individual event, condition or
characteristic required by a given sufficient cause.

Rothman also provided a schematic for these component

causes which have come to be known as causal pies
There has been some work relating sufficient component
causes to potential outcomes (Greenland and Poole 1988,
Rothman and Greenland 1998, Aickin 2002, Flanders 2006, VanderWeele
and Hernan 2006, VanderWeele and Robins 2007, 2008)

Similar ideas also appeared earlier in the philosophical

literature (Mackie, 1965)

E1
G2

G3

Mechanistic Interaction
We may want to know whether two causes G
and E are ever both present in the same
sufficient cause

???

E
G

Negative multiplicative interaction

with no mechanistic interaction
Suppose there are just three mechanisms for outcome D
One requiring G and some other factors A1
One requiring E and some other factors A2
One requiring neither G nor E, just some other factors A0
Suppose that G and E are independent in the population
There is no mechanism
requiring both G and E

A1G
A0
A2E

There is no interaction
between G and E in a
mechanistic sense

Negative multiplicative interaction

with no mechanistic interaction
Suppose G, E, A1, A2 and A0 are all independent in the population
Suppose G and E occur with probabilities 0.2 and 0.5 respectively
Suppose A1 and A2 each occur with probability 0.015
Suppose A0 occurs with probability 0.005
Suppose there are 10,000 subjects in the population
G
E
Total
G=0 E=0
3980
G=1 E=0
980
G=0 E=1
3920
G=1 E=1
965

Cases RR
20
1
20
4.0
80
4.0
35
7.0

Multiplicative interaction:
RR11 / (RR10 x RR01) =7.0/(4 x 4)
= 0.44 < 1
RERI = 7.0 - 4.0 - 4.0 + 1 = 0

We have a non-zero negative multiplicative interaction but This can arise

without any sort of mechanistic interaction! i.e. no mechanism with both G
and E

Positive multiplicative interaction

with no mechanistic interaction
Suppose there are three mechanisms for outcome D
One requiring G and some other factors A1
One requiring E and some other factors A2
One requiring absence of G (denoted by G) and factors A3
Suppose that G and E are independent in the population
There is no mechanism
requiring both G and E

A1G
A3G
A2E

There is no interaction
between G and E in a
mechanistic sense

Positive multiplicative interaction

with no mechanistic interaction
Suppose G and E are independent and occur with probabilities 0.4 and 0.5 respectively
Suppose distribution of A1, A2 and A3 is as follows
P(A1 =0, A2 =1, A3 =1) = .004
P(A1 =1, A2 =0, A3 =0) = .004
P(A1 =0, A2 =1, A3 =0) = .001
P(A1 =1, A2 =1, A3 =0) = .001
P(A1 =0, A2 =0, A3 =0) = .99

Suppose there were 10,000 subjects in the population

G
E
Total
G=0 E=0 3000
G=1 E=0 2000
G=0 E=1 3000
G=1 E=1 2000

Cases RR
12
1
10 1.25
18 1.50
20 2.50

RR11 / (RR10 x RR01)

= 2.50 / (1.25 x 1.50) = 1.33 > 1
RERI = 2.50 - 1.50 - 1.25 + 1 = 0.75 > 0
We have positive multiplicative and
additive interaction
But there is no mechanistic interaction

Notation and Definitions

When can we conclude that there is a mechanism that requires
both exposures?
Counterfactuals: Let Dge denote the counterfactual outcome for an
individual if, possibly contrary to fact, G is set to g and E is set to e
Causal Interaction: We say that there is causal interaction (or
sufficient cause interaction) if for some individual D11=1 but
D10=D01=0
It can be shown that if such causal interaction is present then
there must be synergism in Rothmans sufficient cause framework
(VanderWeele and Robins, 2008)
i.e. a sufficient cause with both G and E

Mechanistic Interaction
We might ask
Are there individuals who would develop breast cancer with the
XRCC3-T241M risk allele and alcohol consumption but not if only
one or the other were present?
Are there individuals who would have diarrheal disease if infected
with both E. coli/Shigella and rotavirus but not if just infected with
one or the other?
Are there individuals who would develop esophageal cancer only if
both of two genetic variants are present?
These are questions about mechanistic interaction, not statistical
interaction

Mechanistic Interaction
Such sufficient cause interaction is not equivalent to statistical
interaction (Greenland and Poole, 1988; Rothman and Greenland,
1998)
Testing for such sufficient cause interaction in general requires
stronger assumptions than statistical interaction
Monotonicity: We will then say that G has a positive monotonic
effect on the outcome D if Dge is non-decreasing in g (similarly for E)
Monotonicity requires the effect always operates in the same direction
for all individuals; it might be plausible sometimes (e.g. the effect of
smoking on lung cancer) but not others (e.g. alcohol on stroke)
Unconfoundedness: We say that the effects of G and E on D are
unconfounded if P(Dge=1)=P(D=1|G=g,E=e)

Mechanistic Interaction
Let pge = P(D=1|G=g,E=e)
Rothman and Greenland (1998) show that if the effects of G and E
on D are unconfounded and if both G and E have positive
monotonic effects on the outcome then one can test for a sufficient
cause interaction by testing:
p11 - p10 - p01 + p00 > 0
i.e. positive additive interaction [the effects of both exposures
combined exceed the sum of the effects of each considered
separately]
We could also test this by RERI > 0
Rothman and Greenland (1998) go on to claim that without
monotonicity conclusions about sufficient cause interaction cannot
be drawn empirically with data

Mechanistic Interaction
Result (VanderWeele and Robins, 2007, 2008): If the effects of G1
and G2 on D are unconfounded one can test for a sufficient cause
interaction by testing:
p11 - p10 - p01 > 0
This condition can be expressed as RERI > 1
It is a stronger condition than simply having positive additive
interaction which would only require RERI > 0
By using this stronger condition one can in fact test for sufficient
cause interaction even without monotonicity, contrary to what was
previously thought
In the 3rd Edition of Modern Epidemiology, Rothman et al. (2008)
correct the claim and discuss these results

Application 1: Breast Cancer

Figueiredo et al. (2004) studied the effects of XRCC3-T241M polymorphisms and
various environmental factors on breast cancer risk
T/T or T/M
M/M

No Alcohol
1.00
1.21 (0.70-2.09)

Alcohol
1.12 (0.81-1.54)
2.09 (1.16-3.78)

Here RERI OR11c - OR10c - OR01c + 1

= 2.09 - 1.21 - 1.12 + 1.00 = 0.76 > 0
The estimate suggests evidence for sufficient cause interaction with the assumption
that both alcohol and the M/M polymorphism have monotonic effects on the
outcome since RERI>0 but that we cannot draw conclusions without monotonicity
since RERI<1
Here, however, it is not very clear whether monotonicity will hold
Moreover, without access to the data we cannot calculate standard errors here

Application 2: Diarrheal Disease

Data from a case-control study in Northwestern Ecuador (2003-2008)
indicates that (Bhavnani et al., 2012):
Giardia
rotavirus
E. coli/Shigella
are all associated with increased risk of diarrheal disease
There is also strong evidence of interaction:
Giardia & rotavirus:
RERI=10.7-2.6-1.1+1 = 7.9 (95% CI: 3.1, 18.9)
E. coli/Shigella & rotavirus: RERI=13.2-2.6-1.6+1 = 9.9 (95% CI: 2.6, 28.4)
E. coli/Shigella & giardia: RERI= 3.0-1.1-1.6+1 = 1.2 (95% CI: -1.4, 3.1)
For Giardia & rotavirus and for E. coli/Shigella & rotavirus, there is strong
evidence of mechanistic interaction even without making any monotoncity
assumptions (VanderWeele, 2012)

Epistasis
Somewhat related ideas and issues appear in the genetics
literature
Often the term epistasis is used to describe statistical interaction
between two genetic factors
Cordell (2002, 2009) points out that although the word epistasis is
now essentially used simply to describe a statistical gene-gene
interaction, the word originally had a somewhat different sense
Bateson (1909) used epistasis to describe instances in which the
effect of a particular genetic variant was masked by a variant at
another locus so that variation of phenotype with genotype at one
locus was only apparent amongst those with certain genotypes at
the second locus

Epistasis
Suppose then we think of interaction this way (different from the statistical
approach) and ask the question whether there are any individuals for
whom the first genetic factor G1 has no effect on the outcome unless the
second genetic factor G2 is present (e.g. G2=1) as in the Table below:
G1=0
G1=1

G2=0
0
0

G2=1
0
1

We would then say that G1 is epistatic to G2 (the effect G1 can be masked

if G2=0)
In our counterfactual notation, this would be: D11=1 but D10=D01=D00=0
Although this is a conceptually distinct notion from that of a statistical
interaction, current terminology does not distinguish between statistical
gene-gene interaction and epistasis in the sense of Bateson (1909)

Epistasis
Cordell (2009) notes that Fisher (1933) used the term epistacy for
statistical gene-gene interaction, distinguishing it from Batesons epistasis
However, the two terms were very similar and with time epistasis came to
be used synonymously with statistical interaction between genetic factors
With the greater recognition that the two are distinct concepts, Phillips
(2008) proposed using statistical epistasis for statistical interaction
between two genetic factors and compositional epistasis for epistasis in
the sense of masking (and functional epistasis for physical interaction)
The terminology has been adopted by others (Cordell, 2009; Moore &
Williams, 2009)
Compositional epistasis: We say that there is compositional epistasis if for
some individual D11=1 but D10=D01=D00=0
(This is stronger than a sufficient cause interaction)

Epistasis
Cordell (2002, 2009) moreover pointed out that tests for statistical
interactions (statistical epistasis) will generally be of limited use in
drawing conclusions about epitasis in the sense of masking
(compositional epistasis) as Bateson had originally conceived of it
Although tests for ordinary statistical interaction between two genetic
factors do not in general allow one to draw conclusions about epistasis,
progress can be made for empirically testing for compositional epistasis
There are relations between empirical data patterns and compositional
epistasis that have not been previously noted and that can be used to
derive non-standard interaction tests to empirically test for such
compositional epistasis (VanderWeele, 2010)

Epistasis
Result (VanderWeele, 2010): If the effects of G1 and G2 on D are
unconfounded then compositional epistasis is present if:
p11 - p10 - p01 - p00 > 0
If at least one of the effects of G1 and G2 are monotonic then:
p11 - p10 - p01 > 0
suffices
If the effects of both G1 and G2 are monotonic then:
p11 - p10 - p01 + p00 > 0
suffices.

Only the final condition (under monotonicity) is the same as an additive

interaction
But we can test for compositional epistasis without monotonicity, using the
other stronger empirical conditions
Results were reported in Nature Reviews Genetics (VanderWeele, 2010)

Epistasis
The measure RERI can also be used for testing for compositional epistasis
RERI = OR11 - OR10 - OR01 + 1 RR11 - RR10 - RR01 + 1
The results above imply that to test for compositional epistasis:
If both G1 and G2 have monotonic effects then we can test:
RERI > 0
If only one of the factors has a monotonic effect we can test:
RERI > 1
Without any monotonicity assumptions we can test:
RERI > 2
The empirical conditions have analogues for multiplicative models
Suppose the outcome is rare and we use a logistic model:
logit {P(D=1|G1=g1,G2=g2)} = 0 + 1g1 + 2g2 + 3g1g2
If main effects 1 and 2 are non-negative the following conditions suffice:
If both G1 and G2 have monotonic effects then we can test:
3 > 0
If only one of the factors has a monotonic effect we can test:
3 > log(2)
Without any monotonicity assumptions we can test:
3 > log(3)

Application 3: Esophageal Cancer

Yang et al. (2005) examine interaction in the effects of Arg variants on ADH2
(chr 4) and Glu/Glu versus Glu/Lys on ALDH2 (chr 12) on esophageal
cancer
Using the case-control data from Yang et al. (2005) to examine additive
interaction using the relative excess risk due to interaction one finds:
RERI = OR11 - OR10 - OR01 + 1 = 7.20 - 1.40 - 3.52 + 1
= 3.28 (95% CI: 0.4, 6.16)
The estimate RERI = 3.28 > 2 would suggest compositional epistasis without
any assumptions about monotonicity at all
The confidence interval contains a value as small as RERI = 0.4 > 0 which
would imply compositional epistasis only if both variants had monotonic
effects on esophageal cancer

Further Remarks
In a recent review article on gene-gene interaction (epistasis)
Phillips (2008) distinguished three types of epistasis:
(1) Statistical epistasis (i.e. interaction in a statistical model)
(2) Compositional epistasis (e.g. D occurs if and only if G1=G2=1)
(3) Functional epistasis (e.g. the physical interaction of proteins)
We have considered new tests for compositional epistasis
It was previously thought that such epistasis could not be detected
using statistical tests (Cordell, 2002); one can test for it but this
requires non-standard interaction test (VanderWeele 2010ab)
But even compositional epistasis does not necessarily imply
functional epistasis, i.e. the physical interaction of proteins

Further Remarks
Suppose that G1 and G2 are two genetic factors
Suppose that when G1=1 protein 1 is not produced
Suppose that when G2=1 protein 2 is not produced
Suppose that the outcome D occurs if and only if neither protein 1
nor protein 2 are present
We then have an epistatic interaction: the outcome occurs if and
only if G1=1 and G2=1
But we do not have physical interaction here
It is precisely the absence of the proteins that gives rise to the
outcome (there is nothing to physically interact here)
It is important to understand the limits of the conclusions being
drawn about these alternative forms of causal interaction

Further Remarks
Sufficient cause interaction was sometimes earlier referred to as biologic
interaction (e.g. Rothman and Greenland, 1998); and sometimes just additive
interaction was even referred to as biologic interaction (Andersson et al., 2005)
As we have seen, neither statistical interaction nor even sufficient cause
interaction necessarily tells us anything about physical or functional interactions
Statistical analyses can only tell us limited information about the underlying
biology (Siemiatycki and Thomas, 1981; Thomas, 1991; Rothman and
Greenland, 1998; Cordell, 2002)

Because of this there has been a suggestion to move away from the use
biologic interaction for sufficient cause interactions (cf. Lawlor, 2011;
VanderWeele, 2011)
It may be more appropriate to refer to these sufficient cause or epistatic
interactions as mechanistic interactions (both exposures together turns the
outcome on and the removal of one turns the outcome off)

Generalizations and Extension

Other subsequent methodological work has generalized these
tests and results to:
Categorical and ordinal exposures (VanderWeele, 2010)
Attributable fractions for interactions (VanderWeele, 2010)
Forms of antagonism (VanderWeele and Knol, 2011)
Time-to-event outcomes (VanderWeele, 2011)
Continuous exposures under dichotomization (VanderWeele et al., 2011;
Berzuini and Dawid, 2012)
n-way interactions (VanderWeele and Richardson, 2012)
Stochastic sufficient causes (VanderWeele and Robins, 2012)
Sensitivity analysis for interaction (VanderWeele et al., 2012)
Power and sample size calculations for sufficient causes interaction
(VanderWeele, 2012)

Application 4: Skin Lesions

High levels of arsenic has been found in well water in Bangladesh
HEALS study was design to assess long-term health consequences of
arsenic and move participants to safer wells
Data from large cohort study in Bangladesh (about 10,000 individuals)
Let X1 for level of arsenic in well water (g/L), a continuous exposure
Let X2=1 for smoking (ever vs. never)
Let D denote our outcome, pre-malignant skin lesions
Covariate control for: sex, age, education, BMI, land and TV ownership
(markers of socioeconomic status in Bangladesh), fertilizer use and
pesticide use
Using methods for continuous exposures, there is evidence for
mechanistic interaction between smoking and arsenic in well water for
arsenic levels above about 80 g/L

Application 4: Skin Lesions

Plot of magnitude of additive interaction by level of well-water arsenic

There is strong evidence for causal interaction around values >80 g/L

Concluding Remarks
(1) Assessing interaction can be important when it is thought that the effect
of one exposure depends on another
(2) When interaction is of interest, both additive and multiplicative interaction
can and should be reported
Additive interaction is always relevant for public health purposes
(3) Mechanistic forms of interaction (sufficient cause and epistatic
interaction) are distinct from statistical interaction
(4) One can empirically test for these with data; the conditions are closely
related to additive interaction
However in each case, without further assumptions about
monotonicity the conditions for these causal interactions are stronger
than simply statistical interactions
(5) It is important to understand the limits of the conclusions being drawn
about these alternative forms of causal interaction

References Part I
Blot WJ, Day NE. Synergism and interaction: are they equivalent? Am. J.
Epidemiol. 1979;110:99-100.
Figueiredo JC, Knight JA, Briollais L, Andrulis IL, Ozcelik H. (2004).
Polymorphisms XRCC1-R399Q and XRCC3-T241M and the risk of breast cancer
at the Ontario Site of the Breast Cancer Family Registry. Cancer Epidemiology,
Biomarkers and Prevention 13:583-591.
Hosmer, D.W., Lemeshow, S. (1992). Confidence interval estimation of
interaction. Epidemiology 3:452-56.
Knol MJ, Egger M, Scott P, Geerlings MI, Vandenbroucke JP. When One
Depends on the Other: Reporting of Interaction in Case-Control and Cohort
Studies. Epidemiology. 2009; 20:161-166.
Knol, M.J. and VanderWeele, T.J. (2012). Guidelines for presenting analyses of
effect modification and interaction. International Journal of Epidemiology,
41:514-520.
Hunter DJ. (2005). Gene-environment interactions in human diseases. Nature
Reviews Genetics, 6:287-298.

References
Lundberg, M., Fredlund, P., Hallqvist, J., Diderichsen, F. (1996). A SAS program
calculating three measures of interaction with confidence intervals. Epidemiology
7:655-656.
Rothman KJ. (1976). Causes. Am J of Epidemiol 104:587-592.
Rothman, K. J. Modern Epidemiology. 1st ed. Little, Brown and Company, Boston,
MA (1986).
Rothman KJ, Greenland S. Modern Epidemiology. Philadelphia: Lippincott-Raven,
1998.
Rothman KJ, Greenland S, Walker AM. Concepts of interaction. Am. J. Epidemiol.
1980;112:467-470.
Saracci R. Interaction and synergism. Am. J. Epidemiol. 1980;112:465-466.

References
Siemiatycki J, Thomas DC (1981). Biological models and statistical interactions: an
example from multistage carcinogenesis. Int. J. Epidemiol. 10:383-387.
VanderWeele, T.J. (2009). On the distinction between interaction and effect
modification. Epidemiology, 20:863-871.
VanderWeele, T.J. and Knol, M.J. (2011). The interpretation of subgroup analyses in
randomized trials: heterogeneity versus secondary interventions. Annals of Internal
Medicine, 154:680-683.
VanderWeele TJ, Robins JM. (2007) The identification of synergism in the SCC
framework. Epidemiol, 18:329-339.
VanderWeele, T.J. and Robins, J.M. (2007). Four types of effect modification a
classification based on directed acyclic graphs. Epidemiology 18:561-568.

References Part II
Andersson, T., Alfredsson, L., Kallberg, H., Zdravkovic, S. and Ahlbom, A. (2005).
Calculating measures of biological interaction. European Journal of Epidemiology
20:575-579.
Cordell, H.J. (2002) Epistasis: what it means, what it doesnt mean, and statistical
methods to detect it in humans. Human Molecular Genetics, 11:2463-2468.
Cordell, H.J. (2009). Detecting gene-gene interaction that underlie human diseases.
Nature Reviews Genetics, 10:392-404.
Greenland, S. and Poole, C. (1988). Invariants and noninvariants in the concept of
interdependent effects. Scandinavian Journal of Work, Environment and Health,
14:125-129.
Lawlor, D.A. (2011). Biological interaction: time to drop the term? Epidemiology,
22:148-50.
Phillips, P.C. (2008). Epistasis the essential role of gene interactions in the
structure and evolution of genetic systems. Nature Reviews Genetic, 9:855-867.
Rothman, K.J. (1976). Causes. American Journal of Epidemiology, 104:587-592.

References
Stern MC, Johnson LR, Bell DA, Taylor JA. XPD codon 751 polymorphism, metabolism
genes, smoking, and bladder cancer risk. Cancer Epidemiology, Biomarkers and Prevention
2002; 11:1004-1011.

Stern MC, Umbach DM, Lunn RM, Taylor JA. DNA repair gene XCRR3 codon 241
polymorphism, its interaction with smoking and XRCC1 polymorphisms, and bladder
cancer risk. Cancer Epidemiology, Biomarkers and Prevention 2002; 11:939-943.
Thomas, W. (1991) Effect modification and the limits of biological inference from
epidemiologic data. Journal of Clinical Epidemiology, 44:221-232.
VanderWeele, T.J. (2009). Sufficient cause interactions and statistical interactions.
Epidemiology, 20:6-13.
VanderWeele, T.J. (2010). Empirical tests for compositional epistasis. Nature Reviews
Genetics, 11:166.
VanderWeele, T.J. (2010). Empirical tests for compositional epistasis. Nature Reviews
Genetics, 11:166.

References
VanderWeele, T.J. (2010). Epistatic interactions. Statistical Applications in Genetics and
Molecular Biology, 9, Article 1:1-22.
VanderWeele, T.J. (2011). A word and that to which it once referred: assessing "biologic"
interaction. Epidemiology, 22:612-613.
VanderWeele, T.J., Hernndez-Diaz, S. and Hernn, M.A. (2010). Case-only geneenvironment interaction studies: when does association imply mechanistic interaction?
Genetic Epidemiology, 34:327-334
VanderWeele, T.J. and Knol, M.J. (2011). Remarks on antagonism. American Journal of
Epidemiology, 173:1140-1147.
VanderWeele, T.J. and Robins J.M. (2007), The identification of synergism in the sufficientcomponent cause framework. Epidemiology, 18:329-339.
VanderWeele, T.J. and Robins, J.M. (2008). Empirical and counterfactual conditions for
sufficient cause interactions. Biometrika, 95:49-61.

References
VanderWeele, T.J., Vansteelandt, S. and Robins, J.M. (2010). Marginal structural models
for sufficient cause interactions. American Journal of Epidemiology, 171:506-514.
Vansteelandt, S., VanderWeele, T.J., Tchetgen, E.J., Robins, J.M., (2008). Multiply robust
inference for statistical interactions. Journal of the American Statistical Association,
103:1693-1704.
Vansteelandt, S., VanderWeele, T.J. and Robins, J.M., Semiparametric inference for
sufficient cause interactions. Journal of the Royal Statistical Society, Series B, in press.
Xu WH, Dai Q, Xiang YB, Long JR, Ruan ZX, Cheng JR, Zheng W, Shu XO. Interaction of
soy food and tea consumption with CYP19A1 Genetic polymorphisms in the development
of endometrial cancer. American Journal of Epidemiology 2007;166:1420-1430.